String-based graders in AI coding agent evals often pass for the wrong reasons. They only confirm a string exists, not the intended behavior. A comment or unused code can trigger a false pass. Absence of a string does not prove absence of the behavior. The author proposes a test: state what a passing or failing grader proves. Simple graders risk false assurance because they are fast and cheap. The solution is to use tools that directly verify the property in question. Build code to check if it compiles, run tests to see if they pass. LLM judges can handle semantic checks but cannot replace direct verification. Agentic evals should be treated like integration tests, not unit tests. Separate gates for building, testing, running and deploying provide clearer evidence. Deterministic graders are useful when the property is deterministic. The core issue is expanding claims beyond what the evidence supports.